Python API Reference

skill-reach provides a programmatic Python API for embedding reachability measurement, manifest linting, and automated description optimization into custom tools or workflows.

---

Top-Level Package (reach)

The most common models and entrypoints are re-exported directly from the top-level reach package:

from pathlib import Path
from reach import lint_tree, load_skills

skills = load_skills(Path("./skills"))
report = lint_tree(Path("./skills"))

for issue in report.issues:
    print(f"[{issue.rule}] {issue.skill}: {issue.message}")

---

Module Index

Module	Description
reach.artifact	Evaluation artifact schema, scoring metrics, and confusion matrix data structures.
reach.catalog	Skill loading, catalog assembly, and YAML frontmatter parsing.
reach.check	Two-stage CI/CD quality gate orchestration, git diff scoping, and threshold enforcement.
reach.cluster	Modularity-based skill clustering for subagent catalog scoping.
reach.config	Configuration loading and Pydantic settings models.
reach.diff	A/B evaluation comparison, noise floor estimation, and effect size reporting.
reach.lint	Static analysis engine for validating skill manifests and listing budgets.
reach.metrics	Accuracy, precision, recall, F1, and multi-step trajectory evaluation metrics.
reach.models	Domain representations: Skill, Catalog, Query, and ProbeResult.
reach.optimize	Closed-loop empirical optimization routines for skill descriptions.
reach.overlap	BM25 vocabulary competition ranking, nearest-rival extraction, and pairwise collision detection.
reach.queries	Labeled evaluation query set loading, serialization, provenance recording, and digesting.
reach.registry	Google Cloud Agent Registry client and skill manifest models.
reach.retrieval	BM25 lexical, dense semantic, and hybrid Reciprocal Rank Fusion (RRF) scorers.
reach.run	Execution engine for planning and conducting evaluation runs.
reach.runtime	Agent execution runtime interfaces, selection outcomes, and driver options models.
reach.sweep	Multi-scale catalog scaling sweeps, capacity knee detection, and loss decomposition.
reach.uncertainty	Wilson score confidence intervals, sample size sizing, and power analysis.